Lecture Notes on

Functional System Reliability Analysis:

An Algorithmic Approach
In using this material, please read and cite:
Abdullah Alkaff and Mochamad Nur Qomarudin, Modeling and Analysis of
System Reliability Using Phase-Type Distribution Closure Properties, Applied
Stochastic Models in Business and Industries, 2020; 1– 22
https://doi.org/10.1002/asmb.2509
Download MATLAB Script:
https://github.com/mochamadnurq/reliability-basicsystem-phasetype
Contents
1. Introduction
1. Background
2. Difficulty in using scalar reliability function
3. Advantages using matrix reliability function
4. Objectives
2. Phase-Type (PH) Distribution
1. Continuous Time Markov Chain (CTMC): definition and modeling
2. Continuous Time Markov Chain (CTMC) with one absorbing state
3. Definition of PH distribution
4. Mathematical model of PH distribution
3. Modeling Component Reliability in PH Distribution
1. Reliability interpretation of PH distribution
2. Reliability measures in PH distribution
3. Special cases of PH distribution
4. Representing other distributions as PH distribution
5. Fitting into PH distribution
Contents (continued)
4. PH Distribution Closure properties (with their proofs)
1. Closure property: definition
2. Kronecker operations
3. Some PH distribution closure properties
1. Mixture operation
2. Addition operation
3. Minimum operation
4. Maximum operation
5. Systems with Basic Structures
1. System reliability modelling approach
2. Series systems
3. Parallel systems
4. K-out-of-N systems
5. Standby systems
6. Standby systems with Imperfect Switch
7. Complexity comparison
Contents (continued)
6. Systems with General Structures
1. Modeling system reliability
2. Algorithm for system reliability analysis
3. Calculating systems reliability measures
4. Calculating confidence interval
7. Discussion
1. Introduction
Contents
1. Motivation
2. Difficulties in using scalar reliability function
3. Advantages of using matrix reliability function
4. Objectives
Motivation
• Conventional computerized algorithmic methods in system reliability analysis: value
based
𝑅1 , 𝑅2 , … , 𝑅𝑁 System 𝑅𝑆
Structure
where:
- 𝑅1 , 𝑅2 , … , 𝑅𝑁 are components’ reliability values
- 𝑅𝑆 is system reliability value
• Ideal computerized algorithmic method in system reliability analysis: functional
based
𝑅1 (𝑡), 𝑅2 (𝑡), … , 𝑅𝑁 (𝑡) System 𝑅𝑆 (𝑡)
Structure
where:
- 𝑅1 (𝑡), 𝑅2 , (𝑡) … , 𝑅𝑁 (𝑡) are components’ reliability functions
- 𝑅𝑆 (𝑡) is system reliability function
Formulas for functional system reliability analysis
No Structure Formula
𝑁
1 Series 𝑅𝑠 (𝑡) = ෑ 𝑅𝑖 (𝑡)
𝑖=1
𝑁
2 Parallel 𝑅𝑠 (𝑡) = 1 − ෑ(1 − 𝑅𝑖 (𝑡))
𝑖=1
𝑁
𝑟𝑖 1−𝑟𝑖
3 K-out-of-N 𝑅𝑠 𝑡 = 1 − ෍ ෑ 𝑅𝑖 𝑡 1 − 𝑅𝑖 𝑡
𝑟1 ,𝑟2 ,…,𝑟𝑛 =0 𝑖=1
𝑟1 +𝑟2 +⋯+𝑟𝑛 ≤𝐾−1
Standby,
𝑅𝑠 𝑡 = 𝑅1 𝑡 ∗ 𝑅2 𝑡 ∗ ⋯ ∗ 𝑅𝑁 (𝑡)
4 Perfect
*: convolution
Switch
𝑅𝑠 𝑡 = 1 − 𝑝1 𝑅1 𝑡 + 𝑝1 1 − 𝑝2 𝑅1 𝑡 ∗ 𝑅2 𝑡 + ⋯
Standby,
𝑁−2 𝑁−1
5 imperfect
+ ෑ 𝑝𝑗 1 − 𝑝𝑁−1 𝑅1 𝑡 ∗ 𝑅2 𝑡 ∗ ⋯ ∗ 𝑅𝑁−1 𝑡 + ෑ 𝑝𝑗 𝑅1 𝑡 ∗ 𝑅2 𝑡 ∗ ⋯ ∗ 𝑅𝑁 𝑡
Switch
𝑗=1 𝑗=1
Difficulties in using scalar reliability function
Let 𝑅𝑖 𝑡 = 𝑒 −𝜆𝑖𝑡 + 𝜆𝑖 𝑡𝑒 −𝜆𝑖 𝑡 = 𝐸𝑟 (2, 𝜆𝑖 )
1 3-Series 𝑅𝑠 (𝑡) = 𝑒 −𝜆1 𝑡 + 𝜆1 𝑡𝑒 −𝜆1 𝑡 𝑒 −𝜆2 𝑡 + 𝜆2 𝑡𝑒 −𝜆2 𝑡 𝑒 −𝜆3 𝑡 + 𝜆3 𝑡𝑒 −𝜆3 𝑡

2 3-Parallel 𝑅𝑠 (𝑡) = 1 − 1 − 𝑒 −𝜆1 𝑡 + 𝜆1 𝑡𝑒 −𝜆1 𝑡 1 − 𝑒 −𝜆2 𝑡 + 𝜆2 𝑡𝑒 −𝜆2 𝑡 1 − 𝑒 −𝜆3 𝑡 + 𝜆3 𝑡𝑒 −𝜆3 𝑡

𝑅𝑠 𝑡 = 𝑒 −𝜆1 𝑡 + 𝜆1 𝑡𝑒 −𝜆1 𝑡 𝑒 −𝜆2 𝑡 + 𝜆2 𝑡𝑒 −𝜆2 𝑡 𝑒 −𝜆3 𝑡 + 𝜆3 𝑡𝑒 −𝜆3 𝑡
+ 𝑒 −𝜆1 𝑡 + 𝜆1 𝑡𝑒 −𝜆1 𝑡 𝑒 −𝜆2 𝑡 + 𝜆2 𝑡𝑒 −𝜆2 𝑡 1 − 𝑒 −𝜆3 𝑡 − 𝜆3 𝑡𝑒 −𝜆3 𝑡
3 2-out-of-3
+ 𝑒 −𝜆1 𝑡 + 𝜆1 𝑡𝑒 −𝜆1 𝑡 1 − 𝑒 −𝜆2 𝑡 − 𝜆2 𝑡𝑒 −𝜆2 𝑡 𝑒 −𝜆3 𝑡 + 𝜆3 𝑡𝑒 −𝜆3 𝑡
+ 1 − 𝑒 −𝜆1 𝑡 − 𝜆1 𝑡𝑒 −𝜆1 𝑡 𝑒 −𝜆2 𝑡 + 𝜆2 𝑡𝑒 −𝜆2 𝑡 𝑒 −𝜆3 𝑡 + 𝜆3 𝑡𝑒 −𝜆3 𝑡
3-Standby
𝑅𝑠 𝑡 = 𝑒 −𝜆1 𝑡 + 𝜆1 𝑡𝑒 −𝜆1 𝑡 ∗ 𝑒 −𝜆2 𝑡 + 𝜆2 𝑡𝑒 −𝜆2 𝑡 ∗ 𝑒 −𝜆3 𝑡 + 𝜆3 𝑡𝑒 −𝜆3 𝑡 → complicated
4 Perfect
*: convolution
Switch
3-Standby
5 Imperfect 𝑅𝑠 𝑡 = 1 − 𝑝1 𝑅1 𝑡 + 𝑝1 1 − 𝑝2 𝑅1 𝑡 ∗ 𝑅2 𝑡 + 𝑝1 𝑝2 𝑅1 𝑡 ∗ 𝑅2 𝑡 ∗ 𝑅3 (𝑡)
Switch

Can be handled using symbolic processing, but the input and outputs are complicated
Difficulties in using scalar reliability function
• How about finding the system hazard function? Must use differentiation
𝑅 ′ 𝑡
ℎ 𝑡 =− ൘𝑅 𝑡
For 3-series system, the simplest among the above mentioned structures:
𝑑
𝑒 −𝜆1 𝑡 + 𝜆1 𝑡𝑒 −𝜆1 𝑡 𝑒 −𝜆2 𝑡 + 𝜆2 𝑡𝑒 −𝜆2 𝑡 𝑒 −𝜆3 𝑡 + 𝜆3 𝑡𝑒 −𝜆3 𝑡
ℎ𝑆 𝑡 = − 𝑑𝑡 −𝜆 𝑡
𝑒 1 + 𝜆1 𝑡𝑒 −𝜆1 𝑡 𝑒 −𝜆2 𝑡 + 𝜆2 𝑡𝑒 −𝜆2 𝑡 𝑒 −𝜆3 𝑡 + 𝜆3 𝑡𝑒 −𝜆3 𝑡
Again, can be handled using symbolic processing, but …
• How about finding the MTTF and VTTF? Must use integration
∞ ∞
MTTFS = න 𝑅𝑆 (𝑡)𝑑𝑡 VTTF = 2 න 𝑡𝑅𝑆 (𝑡)𝑑𝑡 − (MTTFS )2
0 0
For 3-series system, the simplest among the above mentioned structures:
∞
MTTFS = න 𝑒 −𝜆1 𝑡 + 𝜆1 𝑡𝑒 −𝜆1 𝑡 𝑒 −𝜆2 𝑡 + 𝜆2 𝑡𝑒 −𝜆2 𝑡 𝑒 −𝜆3 𝑡 + 𝜆3 𝑡𝑒 −𝜆3 𝑡 𝑑𝑡
0
Again, can be handled using numerical integration, but …
Using matrix reliability function
• The two-stage Erlang scalar reliability function,
𝑅𝑖 𝑡 = 𝑒 −𝜆𝑖 𝑡 + 𝜆𝑖 𝑡𝑒 −𝜆𝑖 𝑡
can be written in matrix reliability function as
𝑅𝑖 𝑡 = 𝒂𝑒 𝑨𝑖 𝑡 𝒖
−𝜆𝑖 𝜆𝑖 1
where 𝒂 = 1 0 , 𝑨𝑖 = ,𝒖 =
0 −𝜆𝑖 1
• The result follows from Jordan canonical form
−𝜆𝑖 𝜆𝑖
0 −𝜆𝑖
𝑡 𝑒 −𝜆𝑖 𝑡 𝜆𝑖 𝑡𝑒 −𝜆𝑖𝑡
𝑒 =
0 𝑒 −𝜆𝑖 𝑡
1
Pre-multiplying by 𝒂 = 1 0 and post-multiplying by 𝒖 = will produce the
1
scalar reliability function of two-stage Erlang.
Using matrix reliability function
• Likewise, also applicable for many other reliability functions
• Using this matrix representation, the system reliability function, say for 3-Series
system, can be expressed as
𝑅𝑆 𝑡 = (𝒂1 𝑒 𝑨1 𝑡 𝒖1 )(𝒂2 𝑒 𝑨2 𝑡 𝒖2 )(𝒂3 𝑒 𝑨3 𝑡 𝒖3 )
which after some matrix algebra manipulation will result in
𝑅𝑆 𝑡 = 𝒂𝑆 𝑒 𝑨𝑆𝑡 𝒖𝑺
• Similar results hold for other basic structures
• This simple expression will simplify the process of finding:
The system hazard function → without using differentiation
The system MTTF and VTTF → without using integration
Objectives
• The design of algorithm
• Receives parameters of 𝑅1 (𝑡), 𝑅2 (𝑡), … , 𝑅𝑁 𝑡 as inputs
• Produces parameters of 𝑅𝑆 (𝑡) as outputs
• Use relationship described in system structures
• The approach of the proposed method
• Represent component reliability function as matrix-based phase-type (PH) distribution
• Utilize closure properties of PH distribution
• Use matrix algebra known as Kronecker, or Tensor, Operations
• The advantage is other reliability measures can be obtained directly
• System hazard function
• MTTF and VTTF
• The implementation in computer programs is easy
• Use only algebraic operations
• No symbolic processing
2. PH Distribution
Contents
1. Continuous Time Markov Chain (CTMC): definition and modeling
2. Continuous Time Markov Chain (CTMC) with one absorbing state
3. Definition of PH distribution
4. Mathematical model of PH distribution
Continuous time Markov chain: definition
• Continuous time discrete state stochastic processes that satisfy Markov property
• Time spent in each state is exponentially distributed
• Let 𝑋𝑖 be the time spent at state 𝒊, then 𝑋𝑖 ~𝐸𝑥𝑝 λ𝑖
• Thus, λ𝑖 is the transition rate out of state 𝑖
• Out of state 𝑖, it makes a transition to state 𝒋 ≠ 𝒊 with probability 𝑝𝑖𝑗 ; hence
∑𝑗≠𝑖 𝑝𝑖𝑗 = 1
• Let 𝑞𝑖𝑗 be the transition rate entering state 𝒋 ≠ 𝒊 from state 𝑖, then
𝑞𝑖𝑗 = 𝑝𝑖𝑗 λ𝑖 ≥ 0 → ∑𝑗≠𝑖 𝑞𝑖𝑗 = λ𝑖
• 𝑞𝑖𝑖 is defined as the negative of the transition rate out of state 𝒊; hence
𝑞𝑖𝑖 = −λ𝑖 ≤ 0
𝑞𝑖𝑖 = −∑𝑗≠𝑖 𝑞𝑖𝑗 → ∑𝑗 𝑞𝑖𝑗 = 𝑞𝑖𝑖 + ∑𝑗≠𝑖 𝑞𝑖𝑗 = 0
• A state is an absorbing state if λ𝑖 = 0, otherwise it is a transient state
Continuous time Markov chain: modeling
• Continuous time Markov chain can be modeled using rate transition matrix, as its
infinitesimal generator, defined as
State (1) (2) … (𝑛)
−λ1 𝑝12 λ1 … 𝑝1𝑛 λ1 (1) 𝑞11 𝑞12 … 𝑞1𝑛 Σ = 0
𝑝21 λ2 −λ2 … 𝑝2𝑛 λ2 (2) 𝑞21 𝑞22 … 𝑞2𝑛 Σ = 0
𝑸= = 𝑞𝑖𝑗 =
… … … … (…) … … … … Σ=0
𝑝𝑛1 λ𝑛 𝑝𝑛2 λ𝑛 … −λ𝑛 (𝑛) 𝑞𝑛1 𝑞𝑛2 … 𝑞𝑛𝑛 Σ = 0
• 𝑸 has properties:
• Off-diagonal elements are nonnegative
• Diagonal elements are nonpositive
• Row sum is equal to zero
• Row elements associated with an absorbing state are all zeros
• Continuous time Markov chain requires initial state probability vector 𝑎𝑖 defined as
the probability that the process starts at state 𝑖. Hence,
𝑎𝑖 ≥ 0 ∑𝑖 𝑎𝑖 = 1
Continuous time Markov chain: with one absorbing state
• PH distribution related to Continuous Time Markov Chain, with one absorbing state
• Let states {1, 2, … , 𝑛}: transient states and {𝑛 + 1}: be the absorbing states
• Its infinitesimal generator can be written as
State (1) (2) … (𝑛 + 1)
(1) −λ1 𝑞12 … 𝑞1,𝑛+1 Σ=0
𝑸 = 𝑞𝑖𝑗 = (2) 𝑞21 −λ2 … 𝑞2,𝑛+1 Σ=0
(…) … … … … Σ=0
(𝑛 + 1) 0 0 … 0 Σ=0
which can be rewritten as
𝑨 𝒃
𝑸= where 𝑨𝒖 + 𝒃 = 𝟎 (since 𝑸𝒖 = 𝟎)
𝟎 0
• Matrix 𝑨 represents transitions among transient states (state transition matrix),
and vector 𝒃 represents transitions from any transient states into the absorbing
states (state absorbing vector)
Continuous time Markov chain: rate transition diagram
• Graph for showing transitions from any state to any state
• Node: state
• Arc: transition
• Node weight: transition rate out of the state
• Arc weight: transition rate or probability of making the transition

𝑞𝑖𝑗 = 𝑝𝑖𝑗 λ𝑖 ≥ 0 → Σ𝑗≠𝑖 𝑞𝑖𝑗 = λ𝑖

Continuous time Markov chain: rate transition diagram
• For simplicity, CTMC can be represented by matrices
State (1) (2) (3) (4)
(1) −𝜆1 𝑞12 0 𝑞14 Σ=0
(2) 0 −𝜆2 𝑞23 𝑞24 Σ=0
𝑸 = (3)
0 0 −𝜆3 𝑞34 Σ=0
(4) 0 0 0 0 Σ=0
𝑨 𝒃
𝑸= 𝒂 = 𝑎1 , 0, 𝑎3
𝟎 0
−𝜆1 𝑞12 0 𝑞14
𝑨= 0 −𝜆2 𝑞23 , 𝒃 = 𝑞24
0 0 −𝜆3 𝑞34
−𝜆1 + 𝑞12 + 𝑞14 0
𝑨𝒖 + 𝒃 = −𝜆2 + 𝑞23 + 𝑞24 = 0
−𝜆3 + 𝑞34 0
Phase-type distribution: definition
• A particular interest in analysis of CTMCs with one absorbing state is the time for
the process to enter the absorbing state for the first time and stay there forever
• Also called time to absorption or absorbing time which is a random variable
• Phase-type (PH) distribution is the distribution of time to absorption from any
initial state in such CTMCs
• Let 𝑋 be the time to absorption in such CTMCs with infinitesimal generator
𝑨 𝒃
𝑸=
𝟎 𝟎
and initial state probability vector from transient states
𝒂 = [𝑎(1), 𝑎(2), … , 𝑎(𝑛)]
• Note that 𝑎 𝑛 + 1 , probability of starting from the absorbing state, is not a
member of 𝒂. Hence,
𝒂𝒖 = 1 − 𝑎 𝑛 + 1 ≤ 1
where 𝒖 = 1, 1, … , 1 ′ where ‘ denotes transpose
Mathematical model of PH distribution
• Neuts (1994) showed that 𝑋 has cumulative distribution function
𝐹𝑋 (𝑥) = 1 − 𝒂𝑒 𝑨𝑥 𝒖 𝑥≥0
• 𝐹𝑋 0 = 1 − 𝒂𝑒 𝟎 𝒖 = 𝑎(𝑛 + 1)→ 𝑋 has a discontinuity/jump, at 𝑥 = 0 size 𝑎(𝑛 + 1)
• 𝐹𝑋 (𝑥) = 1 − 𝒂𝑒 𝑨𝑥 𝒖 is said ‘PH distribution with representation (𝒂, 𝑨)’
• 𝑋~𝑃𝐻(𝑛, 𝒂, 𝑨) means ‘𝑋 has a PH distribution of order 𝑛 with representation (𝑎, 𝐴)’
• Its probability density function is the derivative of 𝐹𝑋 (𝑥) and since 𝐹𝑋 (𝑥) has
discontinuity of size 𝑎(𝑛 + 1) at 𝑥 = 0, then
𝑓𝑋 𝑥 = 𝑑𝐹𝑋 𝑥 /𝑑𝑥 = −𝒂𝑨𝑒 𝑨𝑥 𝒖 + 𝑎(𝑛 + 1)δ(𝑥) 𝑥≥0
where δ(𝑥) is an impulse function at 𝑥 = 0
• The 𝑖𝑡ℎ moment of 𝑋
𝐸[𝑋 𝑖 ] = (−1)𝑖 𝑖! 𝒂𝑨−𝑖 𝒖 𝑖 = 1, 2, …
Neuts, MF. (1994). Matrix-geometric solutions in Stochastic systems: an algorithmic approach, Dover, NY.
Mathematical model of PH distribution
Example: A CTMC and PH representation of its time to absorption
𝐹𝑋 𝑥 = 1 − 𝒂𝑒 𝑨𝑥 𝒖 𝑓𝑋 𝑥 = −𝒂𝑨𝑒 𝑨𝑥 𝒖 + 𝑎 𝑛 + 1 δ 𝑥 𝑥≥0
State (1) (2) (3) (4)
(1) −14 4 0 10 Σ=0
(2) 0 −15 6 9 Σ=0
𝑸=
(3) 0 0 −8 8 Σ=0
(4) 0 0 0 0 Σ=0
−14 4 0 10
𝑨= 0 −15 6 ,𝒃 = 9
0 0 −8 8
1 1 1 1
𝒂= ,𝑎 𝑛 + 1 = ,𝑛 = 3
2 5 5 10
Time spent at state 1: exponential (14)
Prob. making transition from state 1 to 2 : 4/14
Mathematical model of PH distribution: example
𝐹𝑋 𝑥 = 1 − 𝑒 −𝜆𝑥 − 𝜆𝑥𝑒 −𝜆𝑥 (2- Erlang) 𝐹𝑋 𝑥 = 1 − 0.8𝑒 −𝜆𝑥 − 0.8𝜆𝑥𝑒 −𝜆𝑥
−𝜆 𝜆 −𝜆 𝜆
𝐹𝑋 𝑥 = 1 − 1, 0 exp 𝑥 𝒖 𝐹𝑋 𝑥 = 1 − 0.8, 0 exp 𝑥 𝒖
0 −𝜆 0 −𝜆
𝒂𝒖 = 1 → 𝑎 𝑛 + 1 = 0 𝒂𝒖 = 0.8 → 𝑎 𝑛 + 1 = 1 − 0.8 = 0.2

𝑎 𝑛+1 =0 𝑎 𝑛 + 1 = 0.2
Mathematical model of PH distribution: example
Acyclic PH (APH) Cyclic PH (CPH)
−0.1 0.1 −0.6 0.5
𝐹𝑋 𝑥 = 1 − 1, 0 exp 𝑥 𝒖 𝐹𝑋 𝑥 = 1 − 0.6, 0.4 exp 𝑥 𝒖
0 −0.2 0.3 −0.5
A is an upper-diagonal matrix −0.6 0.5 0.1
−0.1 0.1 0
𝑸= 0 𝑸 = 0.3 −0.5 0.2
−0.2 0.2
0 0 0 0 0 0
3. Modeling Component Reliability in PH
Distribution
Contents
1. Reliability interpretation of PH distribution
2. Reliability measures in PH distribution
3. Special cases of PH distribution
4. Representing other distributions as PH distribution
5. Fitting into PH distribution
Reliability interpretation of PH distribution
• Component’s condition is degraded over time, caused by fatigues, shocks, etc.
• The condition can be upgraded by maintenance,
• Such degrading/upgrading processes can be represented as a CTMC
• State(s) that represent unacceptable (failed) condition is the absorbing state(s)
• Otherwise, they are transient states
• Example:

• The absorbing (failed) state may represents inoperable, critical, or even good state
• Time to absorption is the same as time to failure or lifetime
Reliability measures in PH distribution
• Reliability function of a component in the PH distribution is
𝑅 𝑡 = 𝑃 𝑇 > 𝑡 = 1 − 𝐹𝑇 𝑡 = 𝒂𝑒 𝑨𝑡 𝒖 + 𝑎 𝑛 + 1 δ 𝑥 𝑡≥0
where 𝑇 is the component’s lifetime.
• In most cases, and assumed in this notes, 𝑎 𝑛 + 1 = 0. Thus,
𝑅 𝑡 = 𝒂𝑒 𝑨𝑡 𝒖 𝑡≥0
𝑎 𝑛 + 1 = 1 − 𝒂𝒖 = 0.2
−𝜆𝑡 −𝜆𝑡
𝑅 𝑡 =𝑒 + 𝜆𝑡𝑒
−𝜆 𝜆 𝑅 𝑡 = 0.8𝑒 −𝜆𝑡 + 0.8𝜆𝑡𝑒 −𝜆𝑡
𝑅 𝑡 = 1, 0 exp 𝑡 𝒖
0 −𝜆 −𝜆 𝜆
𝑅 𝑡 = 0.8, 0 exp 𝑡 𝒖
0 −𝜆
Reliability measures in PH distribution
• Hazard function is given by (by assuming there is no discontinuity at 𝑡 = 0)
𝑅′ 𝑡 𝑓𝑇 𝑡
ℎ 𝑡 =− = = −(𝒂𝑨𝑒 𝑨𝑡 𝒖)(𝒂𝑒 𝑨𝑡 𝒖)−1
𝑅 𝑡 𝑅 𝑡
• Example:
Scalar:
𝑅 𝑡 = 𝑒 −𝜆𝑡 + 𝜆𝑡𝑒 −𝜆𝑡
𝜆𝑡𝑒 −𝜆𝑡
ℎ 𝑡 = −𝜆𝑡
𝑒 + 𝜆𝑡𝑒 −𝜆𝑡
Matrix:
−𝜆 𝜆
𝑡
𝑅 𝑡 = 1, 0 𝑒 0 −𝜆 𝒖

−𝜆 𝜆 −𝜆 𝜆 −1
𝑡 𝑡
ℎ 𝑡 = 𝜆, −𝜆 𝑒 0 −𝜆 𝒖 1, 0 𝑒 0 −𝜆 𝒖

• In this example, the component has increasing failure rate, or is an IFR component
Reliability measures in PH distribution
• Mean time to failure (MTTF) can be obtained as
𝑀𝑇𝑇𝐹 = 𝐸[𝑇] = −𝒂𝑨−1 𝒖
• Variance time to failure (MTTF)
𝑉𝑇𝑇𝐹 = 2𝒂𝑨−2 𝒖 − 𝑀𝑇𝑇𝐹 2
• Coefficient of variation of the time to failure (CVTTF) 𝑪𝑽𝑻𝑻𝑭 = 𝟏 → 𝐂𝐅𝐑
𝑉𝑇𝑇𝐹 𝑪𝑽𝑻𝑻𝑭 > 𝟏 → 𝐃𝐅𝐑
𝐶𝑉𝑇𝑇𝐹 = 𝑪𝑽𝑻𝑻𝑭 < 𝟏 → 𝐈𝐅𝐑
𝑀𝑇𝑇𝐹
• To avoid matrix inversion, introduce 𝜷 = 𝒂𝑨−1 and 𝜸 = 𝒂 𝑨2 −1 = 𝜷𝑨−1 , hence
𝑀𝑇𝑇𝐹 = −𝜷𝒖
𝑉𝑇𝑇𝐹 = 2𝜸𝒖 − 𝑀𝑇𝑇𝐹 2
𝜷 and 𝜸 can be obtained by solving 𝜷𝑨 = 𝒂 and 𝜸𝑨 = 𝜷 using Gaussian elimination
Special cases of PH distribution: exponential
• Lifetime of a component with one-stage operating condition → exponential
Lifetime distribution: 𝐹𝑇 𝑡 = 1 − 𝑒 −𝜆𝑡
Scalar Reliability Function:
𝑅 𝑡 = 1 − 𝐹𝑇 𝑡 = 𝑒 −𝜆𝑡

𝑅 𝑡 = 𝑒 −0.01𝑡

−𝜆 𝜆
𝑸=
0 0
• Matrix Reliability Function: 𝑨 = −𝜆
𝑅 𝑡 = 1 𝑒 −𝜆 𝑡 1 𝑎= 1
𝑛=1
Special cases of PH distribution: Erlang
• Lifetime of a component with n-stage operating condition → n-stage Erlang
Lifetime distribution:
𝑛−1
1
𝐹𝑇 𝑡 = 1 − ෍ 𝜆𝑡 𝑖 𝑒 −𝜆𝑡
𝑖!
𝑖=0

Reliability in scalar: −𝜆 𝜆 … 0 0
𝑛−1
1 0 −𝜆 … 0 0
𝑅 𝑡 = ෍ 𝜆𝑡 𝑖 𝑒 −𝜆𝑡 𝑅 𝑡 = 1 + 0.01𝑡 𝑒 −0.01𝑡 𝑸= ⋮ ⋮ ⋱ ⋮ 0
𝑖! 0 0 … −𝜆 𝜆
𝑖=0 0 0 0 0 0
−𝜆 𝜆 … 0
Reliability in matrix: 𝑨=
0 −𝜆 … 0
⋮ ⋮ ⋱ ⋮
𝑅 𝑡 = 𝒂𝑒 𝑨𝑡 𝒖 0 0 … −𝜆
𝒂 = 1, 0, … , 0
𝑛=𝑛
Special cases PH distribution: exponential and Erlang

No Distribution Category Representation of Reliability Function

Constant Graphic
1 Exponential (𝜆) Failure
Scalar 𝑒 −𝜆𝑡
Rate (CFR)
Matrix 𝑛=1 𝒂= 1 𝑨 = [− 𝜆]

Graphic

𝑛−1
Increasing Scalar 1
n-stage Erlang ෍ 𝜆𝑡 𝑖 𝑒 −𝜆𝑡
2 Failure 𝑖!
𝑖=0
(𝜆)
Rate (IFR)
−𝜆 𝜆 … 0
0 −𝜆 … 0
Matrix 𝑛=𝑛 𝒂= 1 0…0 𝑨=
⋮ ⋮ ⋱ ⋮
0 0 … −𝜆
Special cases PH distribution: hypo-exponential

No Distribution Category Representation of Reliability Function

Graphic

𝑛 −𝜆𝑖 𝑥 𝑛
𝑒 𝜆𝑖
Scalar ෍ with 𝑃𝑖 = ෑ 1−
n-stage Hypo- Increasing 𝑃𝑖 𝜆𝑗
𝑖=1 𝑗=1,𝑗≠𝑖
3 exponential Failure
(𝑝𝑖 , 𝜆𝑖 ) Rate (IFR) 𝑛 = 𝑛 𝒂 = 1 0…0
−𝜆1 𝜆1 … 0 0
Matrix 0 −𝜆2 … 0 0
𝑨= ⋮ ⋮ ⋱ ⋮ ⋮
0 0 … −𝜆𝑛−1 𝜆𝑛−1
0 0 … 0 −𝜆𝑛
Special cases PH distribution: hyper-exponential
No Distribution Category Representation of Reliability Function

Graphic

n-stage
Decreasing
Hyper-
4 Failure
exponential
Rate (DFR) 𝑛
(𝜆𝑖 )
Scalar ෍ 𝑝𝑖 𝑒 −𝜆𝑖𝑡
𝑖=1
−𝜆1 0 … 0
0 −𝜆2 … 0
Matrix 𝑛 = 𝑛 𝒂 = 𝑝1 𝑝2 … 𝑝𝑛 𝑨=
⋮ ⋮ ⋱ ⋮
0 0 … −𝜆𝑛
Special cases PH distribution: three-stage Coxian

No Distribution Category Representation of Reliability Function

Graphic

3-stage
Coxian IFR/DFR/
5 (𝑝, 𝑞, 𝜆2 𝜆3 𝜆2 − 𝜆3 𝑒 −𝜆1𝑡 + 𝜆1 𝜆3 (𝜆3 − 𝜆1 )𝑒 −𝜆2𝑡 + 𝜆1 𝜆2 (𝜆1 − 𝜆2 )𝑒 −𝜆3𝑡
BFR* 𝑝𝑞
𝜆1 − 𝜆2 𝜆1 − 𝜆3 𝜆2 − 𝜆3
Scalar
𝜆1 , 𝜆2 , 𝜆3 ) 𝜆2 𝑒 −𝜆1𝑡 − 𝜆1 𝑒 −𝜆2𝑡
+𝑝 1−𝑞 + 1 − 𝑝 𝑒 −𝜆1𝑡
𝜆1 − 𝜆2
−𝜆1 𝑝𝜆1 0
Matrix 𝑛=3 𝒂= 1 0 0 𝑨= 0 −𝜆2 𝑞𝜆2
0 0 −𝜆3

*BFR: Bathtub failure rate

Special cases PH distribution: three-stage Coxian

Category Parameter
IFR 𝑝 = 0.97561, 𝑞 = 0.8, 𝜆1 = 0.41, 𝜆2 = 0.25, 𝜆3 = 0.1
DFR 𝑝 = 0.8, 𝑞 = 0.8, 𝜆1 = 0.5, 𝜆2 = 0.25, 𝜆3 = 0.01
BFR 𝑝 = 0.8, 𝑞 = 0.95238, 𝜆1 = 0.5, 𝜆2 = 0.21, 𝜆3 = 0.1
Faddy, M. (1995). Phase-type distributions for failure times. Mathematical and Computer Modelling, 22(10–12), p. 63–70.
Representing other distributions as PH Distribution
• Many distributions that commonly used to model component reliability, i.e. Weibull
and lognormal, can be represented as PH distribution.
• Example: Weibull (1.5;0.1)
−0.1𝑡 1.5
Scalar: 𝑅 𝑡 = 𝑒
Matrix : 𝑅 𝑡 = 𝒂𝑒 𝑨𝑡 𝒖, where
−0.216 0.213
𝒏=𝟐 𝒂= 1 0 𝑨=
0 −0.216
−0.270 0.147 0.099
𝒏=𝟑 𝒂= 1 0 0 𝑨= 0 −0.270 0.270
0 0 −0.270
−0.339 0.169 0.009 0.154
0 −0.323 0.323 0
𝒏=𝟒 𝒂= 1 0 0 0 𝑨=
0 0 −0.323 0.323
0 0 0 −0.326
Representing other distributions as PH distribution
PH distribution of Weibull (1.5;0.1)

Distribution Mean cv2 Max Dev

Weibull 9.03 0.46 -
2-APH 9.17 0.50 0.0104
3-APH 9.07 0.48 0.0036
4-APH 9.03 0.46 0.00063
Trivedi and Bobbio (2017) Reliability and Availability Engineering: Modelling, Analysis, and Applications. Cambridge Univ. Press.
Fitting into PH distribution
Some software packages for fitting PH distribution (Okamura and Dohi, 2016):
• BuTools (MATLAB)
• PhFit (Java)
• Empht (C)
• Momentmatching (Mathematica/MATLAB)
• G-FIT (Command Line)
• jPhaseFit (Java Library)
• HyperStar (Java)
• mapfit (R)
PH representation of Weibull (1.5, 0.1) given in previous slide was obtained using PhFit

Okamura and Dohi (2016) PH fitting algorithm and its application to reliability engineering, J. of the Oper. Res. Soc. Japan, 59, 1.
Fitting into PH distribution: lognormal
• Example: Lognormal (𝝁 = 𝟏. 𝟓;𝝈 =0.7)
1 1 ln 𝑡−𝜇
Scalar: 𝑅 𝑡 = − erf
2 2 √2𝜎

Matrix (PH): 𝑅 𝑡 = 𝒂𝑒 𝑨𝑡 𝒖, fitted using BuTools

• BuTools is not as good as PhFit, because:
- BuTools uses mixture Erlang
- PHFit uses Coxian
- Coxian is more flexible than mixture Erlang
2
Distribution Mean cv Max Dev Σ𝝐𝟐
Lognormal 5.7259 0.6326 - -
4-APH 5.6759 0.7049 0.0513 0.0140
5-APH 5.6759 0.6775 0.0362 0.0046
6-APH 5.6759 0.6548 0.0300 0.0042
Fitting into PH distribution: lognormal
• PH representation of n-APH for lognormal (𝝁 = 𝟏. 𝟓;𝝈 =0.7)
−0.1030 0 0 0
0 −0.6065 0.6065 0
𝒏 = 𝟒, 𝒂 = 0.1188, 0.8812, 0, 0 𝑨=
0 0 −0.6065 0.6065
0 0 0 −0.6065
−0.2351 0.2351 0 0 0
0 −0.2351 0 0 0
𝒏 = 𝟓, 𝒂 = 0.2693, 0, 0.7307, 0, 0 𝑨= 0 0 −0.6805 0.6805 0
0 0 0 −0.6805 0.6805
0 0 0 0 −0.6805
−0.3153 0.3153 0 0 0 0
0 −0.3153 0.3153 0 0 0
0 0 −0.3153 0 0 0
𝒏 = 𝟔, 𝒂 = 0.2406, 0, 0, 0.7594, 0, 0 𝑨=
0 0 0 −0.7069 0.7069 0
0 0 0 0 −0.7069 0.7069
0 0 0 0 0 −0.7069
4. PH Distribution Closure Properties
Contents
1. Closure property: definition
2. Kronecker operations
3. Some PH distribution closure properties
1. Mixture Operation
2. Addition Operation
3. Minimum Operation
4. Maximum Operation
Closure property
• Closure property is a property of a set that certain operations to all members of the
set will result in members of the same set
• Example: 𝛦 = exponential function
𝑥 𝑡 = 𝑎𝑒 𝑏𝑡 𝜖 𝛦 → 𝑥ሶ 𝑡 = 𝑎𝑏𝑒 𝑏𝑡 𝜖 𝛦
𝑏𝑡
𝑎 𝑏𝑡
𝑥 𝑡 = 𝑎𝑒 𝜖 𝛦 → න𝑥 𝑡 𝑑𝑡 = 𝑒 𝜖 𝛦
𝑏

(𝑥 𝑡 = 𝑎𝑒 𝑏𝑡 𝜖 𝛦 ∧ 𝑦 𝑡 = 𝑐𝑒 𝑑𝑡 𝜖 𝛦) → 𝑥 𝑡 𝑦 𝑡 = 𝑎𝑐𝑒 𝑏+𝑑 𝑡
𝜖𝛦
𝑥 𝑡 𝑎 𝑏−𝑑 𝑡
(𝑥 𝑡 = 𝑎𝑒 𝑏𝑡 𝜖 𝛦∧𝑦 𝑡 = 𝑐𝑒 𝑑𝑡 𝜖 𝛦) → = 𝑒 𝜖 𝛦
𝑦 𝑡 𝑏

• 𝛦 is closed under differentiation, integration, multiplication, and division operations

Kronecker operations
• PH distribution has several closure properties
• All of them involve matrix algebra operations known as Kronecker operations
comprises of Kronecker product (⊗) and Kronecker sum (⊕)
• For matrices 𝑨 = [𝐴𝑖𝑗 ] and 𝑩 = [𝐵𝑖𝑗 ] of dimensions 𝑚1 × 𝑚2 and 𝑛1 × 𝑛2 ,

𝐴11 𝑩 𝐴12 𝑩 ⋯ 𝐴1𝑚2 𝑩

𝑨⊗𝑩= ⋮ ⋮ ⋱ ⋮ of dimension 𝑚1 𝑛1 × 𝑚2 𝑛2
𝐴𝑚1 1 𝑩 𝐴𝑚1 2 𝑩 ⋯ 𝐴𝑚 1 𝑚2 𝑩

• For 𝑨 and 𝑩 of dimension 𝑚 × 𝑚 and 𝑛 × 𝑛,

𝑨 ⊕ 𝑩 = 𝑨 ⊗ 𝑰𝑛 + 𝑰𝑚 ⊗ 𝑩 of dimension 𝑚𝑛 × 𝑚𝑛

where 𝑰𝑚 and 𝑰𝑛 are identity matrices of order 𝑚 and 𝑛, respectively.

Kronecker operations
• Example:
5 5 5 0 15
𝑨= 0 3 ,𝑩 = →𝑨⊗𝑩 = 0 3 =
−3 −3 −3 0 −9
1 15 1 15 1 15 2 0 3 15
𝑨= ,𝑩 = 2 → 𝑨 ⊕𝑩 = ⊗ 𝐼1 + 𝐼2 ⊗ 2 = + =
3 −9 3 −9 3 −9 0 2 3 −7
• Some properties of Kronecker operations utilized in this study are:
𝑔 ⊗ ℎ = 𝑔ℎ
𝑔 ⊗ 𝑨 = 𝑔𝑨
𝑨𝑩 ⊗ 𝑪𝑫 = 𝑨 ⊗ 𝑪 𝑩 ⊗ 𝑫
𝑒 𝑨𝑡 ⊗ 𝑒 𝑩𝑡 = 𝑒 𝑨⊕𝑩 𝑡

𝒖𝑚 ⊗ 𝒖𝑛 = 𝒖𝑚𝑛
where 𝑔 and ℎ are scalars and 𝑨, 𝑩, 𝑪, 𝑫 are matrices.
Kronecker operations
• Example: Multiplication of two scalar-valued matrix exponential functions
𝑥𝑖 𝑡 = 𝒂𝑖 𝑒 𝑨𝑖𝑡 𝒖𝑖
where 𝒂𝑖 is 1 × 𝑛𝑖 , 𝒖𝑖 is 𝑛𝑖 × 1, 𝑨𝑖 is 𝑛1 × 𝑛2
Hence, 𝑥𝑖 𝑡 is a scalar-valued matrix exponential function

𝑥1 𝑡 𝑥2 𝑡 = 𝒂1 𝑒 𝑨1 𝑡 𝒖1 𝒂2 𝑒 𝑨2 𝑡 𝒖2

= 𝒂1 𝑒 𝑨1 𝑡 𝒖1 ⊗ 𝒂2 𝑒 𝑨2𝑡 𝒖2 (use 𝑔 ⊗ ℎ = 𝑔ℎ)

= 𝒂1 𝑒 𝑨1 𝑡 ⊗ 𝒂2 𝑒 𝑨2 𝑡 𝒖1 ⊗ 𝒖2 (use 𝑨𝑩 ⊗ 𝑪𝑫 = 𝑨 ⊗ 𝑪 𝑩 ⊗ 𝑫 )
= 𝒂1 ⊗ 𝒂2 𝑒 𝑨1𝑡 ⊗ 𝑒 𝑨2𝑡 𝒖1 ⊗ 𝒖2 (use 𝑨𝑩 ⊗ 𝑪𝑫 = 𝑨 ⊗ 𝑪 𝑩 ⊗ 𝑫 )

= 𝒂1 ⊗ 𝒂2 𝑒 𝑨1 ⊕𝑨2 𝑡 𝒖1 ⊗ 𝒖2 (use 𝑒 𝑨𝑡 ⊗ 𝑒 𝑩𝑡 = 𝑒 𝑨⊕𝑩 𝑡 )

Some PH distribution closure properties
• PH distribution is closed for the following operations:
1. Mixture operation
2. Addition operation
3. Minimum operation
4. Maximum operation
• Let
𝑇1 ~𝑃𝐻(𝑛1 , 𝒂1 , 𝑨1 )
𝑇2 ~𝑃𝐻(𝑛2 , 𝒂2 , 𝑨2 )
and they are statistically independent
• Let
𝑇𝑆 ~𝑃𝐻(𝑛𝑆 , 𝒂𝑆 , 𝑨𝑆 )
be the result of those operations to 𝑇1 and 𝑇2
Mixture operation: 𝑇𝑆 = MIX 𝑇1 , 𝑇2 , 𝑝1 , 𝑝2
𝑇𝑆 = MIX 𝑇1 , 𝑇2 , 𝑝1 , 𝑝2 = 𝑇1 with probability 𝑝1
𝑇𝑆
= 𝑇2 with probability 𝑝2
where 𝑝1 + 𝑝2 = 1
Mixture operation: 𝑇𝑆 = MIX 𝑇1 , 𝑇2 , 𝑝1 , 𝑝2
𝐹𝑇𝑆 𝑡 = 𝑝1 𝐹𝑇1 𝑡 + 𝑝2 𝐹𝑇2 𝑡
= 𝑝1 1 − 𝒂1 𝑒 𝑨1 𝑡 𝒖1 + 𝑝2 1 − 𝒂2 𝑒 𝑨2 𝑡 𝒖2 = 1 − 𝑝1 𝒂1 𝑒 𝑨1 𝑡 𝒖1 − 𝑝2 𝒂2 𝑒 𝑨2 𝑡 𝒖2
𝑒 𝑨1 𝑡 𝟎 𝒖1
= 1 − 𝑝1 𝒂1 , 𝑝2 𝒂2
𝟎 𝑒 𝑨2 𝑡 𝒖2
𝑨1 𝟎 𝒖1
= 1 − 𝑝1 𝒂1 , 𝑝2 𝒂2 exp 𝑡 𝒖 = 1 − 𝒂𝑆 𝑒 𝑨𝑆𝑡 𝒖𝑆
𝟎 𝑨2 2

where
𝒂𝑆 = 𝑝1 𝒂1 , 𝑝2 𝒂2
𝑨1 𝟎
𝑨𝑆 =
𝟎 𝑨2
𝒖1
𝒖𝑆 = 𝒖 .
2

From the vectors and matrix involved, the resulted PH distribution has order
𝑛𝑆 = 𝑛1 + 𝑛2
Mixture operation: 𝑇𝑆 = MIX 𝑇1 , 𝑇2 , 𝑝

Theorem 1. (Neuts, 1994)

MIX 𝑇1 ,𝑇2 , 𝑝 has a PH distribution with representation (𝑛𝑆 , 𝒂𝑆 , 𝑨𝑆 ) given by
𝑛𝑆 = 𝑛1 + 𝑛2
𝒂𝑆 = 𝑝1 𝒂1 , 𝑝2 𝒂2
𝑨1 𝟎
𝑨𝑆 =
𝟎 𝑨2
Addition operation: ADD (𝑇1 , 𝑇2 )

𝑇𝑆 = ADD 𝑇1 , 𝑇2 = 𝑇1 + 𝑇2

State (1) (2) (0) 𝑇𝑆

(1) 𝑨1 𝒃1 𝒂2 𝟎
𝑸𝑆 = (2) 𝟎 𝑨2 𝒃2
(0) 𝟎 𝟎 𝟎
𝑨1 𝒃1 𝒂2
𝑨𝑆 =
𝟎 𝑨2
𝒂𝑆 = 𝒂1 , 𝑎1 𝑛1 + 1 𝒂2 = 𝒂1 , 𝟎

The order is 𝑛𝑆 = 𝑛1 + 𝑛2
Addition operation: ADD (𝑇1 , 𝑇2 )
Theorem 2. (Neuts, 1994)
ADD (𝑇1 , 𝑇2 ) has a PH distribution with representation (𝑛𝑠 , 𝒂𝑠 , 𝑨𝑠 ) given by
𝑛𝑠 = 𝑛1 + 𝑛2
𝒂𝑠 = [𝒂1 , 𝑎1 (𝑛1 + 1)𝒂2 ]
𝑨1 𝒃1 𝒂2
𝑨𝑠 =
𝟎 𝑨2
where
𝒃1 = −𝑨1 𝒖

and 𝑎𝑖 𝑛𝑖 + 1 is the probability of starting at the absorbing state

Minimum operation: 𝑇𝑆 = MIN 𝑇1 , 𝑇2
𝐹𝑇𝑆 𝑡 = 𝑃 MIN 𝑇1 , 𝑇2 ≤ 𝑡 = 1 − 𝑃 MIN 𝑇1 , 𝑇2 > 𝑡
= 1 − 𝑃 𝑇1 > 𝑡 ∧ 𝑇2 > 𝑡 = 1 − 𝑃 𝑇1 > 𝑡)𝑃(𝑇2 > 𝑡
= 1 − 1 − 𝐹𝑇1 𝑡 1 − 𝐹𝑇2 𝑡 independent assumption

= 1 − 𝒂1 𝑒 𝑨1 𝑡 𝒖1 𝒂2 𝑒 𝑨2 𝑡 𝒖2
= 1 − 𝒂𝑆 𝑒 𝑨𝑆𝑡 𝒖𝑆 result of previous section
where
𝒂𝑆 = 𝒂1 ⊗ 𝒂2
𝑨𝑆 = 𝑨1 ⊕ 𝑨2
𝒖𝑆 = 𝒖1 ⊗ 𝒖2
Kronecker operations will yield vectors and matrix with order 𝑛𝑆 = 𝑛1 𝑛2
Minimum operation: 𝑇𝑆 = MIN 𝑇1 , 𝑇2

Theorem 3. (Neuts, 1994)

MIN 𝑇1 ,𝑇2 has a PH distribution with representation (𝑛𝑆 , 𝒂𝑆 , 𝑨𝑆 ) given by
𝑛𝑆 = 𝑛1 𝑛2
𝒂𝑆 = 𝒂1 ⊗ 𝒂2
𝑨𝑆 = 𝑨1 ⊕ 𝑨2
Maximum operation: 𝑇𝑆 = MAX 𝑇1 , 𝑇2
• Maximum operation 𝑇𝑆 = MAX 𝑇1 , 𝑇2 implies the largest of 𝑇1 𝑎𝑛𝑑 𝑇2
• Starts with both are in transient states
• Stays there until one of them enters absorbing state, that is, the minimum of the two
(𝑇1 and 𝑇2 ) enters the absorbing state
• There are two possibilities: 𝑇1 or 𝑇2 is the minimum
• Make a transition from both are in transient states to either one is in transient states
• Stays there until the remaining of the two (𝑇1 or 𝑇2 ) enters the absorbing state
• Makes a transition from either one is in transient states to both are in absorbing states
• 𝑇S represents the duration from both are in transient states to both are in absorbing states
Maximum operation: 𝑇𝑆 = MAX 𝑇1 , 𝑇2
𝑇𝑆

𝑇1 : Transient
𝑇2 : Absorbed

Starting 𝑇1 : Absorbed
probability 𝑇2 : Absorbed
𝑇1 : Transient 𝑇1 : Absorbed
𝑇2 : Transient 𝑇2 : Transient
(Until min of
𝑇1 and 𝑇2 )
Maximum operation: 𝑇𝑆 = MAX 𝑇1 , 𝑇2
𝑇1 transient

Kronecker product
informs that the
min 𝑇1 , 𝑇2 transition may occur at
all transient states of 𝑇1
or 𝑇2

𝑇1 , 𝑇2 transient

𝑇2 transient
Maximum operation: 𝑇𝑆 = MAX 𝑇1 , 𝑇2
State (1,1) (1,0) (0,1) (0,0)
(1,1) 𝑨1 ⊕ 𝑨2 𝑰𝑛1 ⊗ 𝒃2 𝒃1 ⊗ 𝑰𝑛2 𝟎
𝑸𝑆 = (1,0) 𝟎 𝑨1 𝟎 𝒃1
(0,1) 𝟎 𝟎 𝑨2 𝒃2
(0,0) 𝟎 𝟎 𝟎 𝟎

𝑨1 ⊕ 𝑨2 𝑰𝑛1 ⊗ 𝒃2 𝒃1 ⊗ 𝑰𝑛2
𝑨𝑆 = 𝟎 𝑨1 𝟎
𝟎 𝟎 𝑨2

𝒂𝑆 = [𝒂1 ⊗ 𝒂2 , 𝑎2 (𝑛2 + 1)𝒂1 , 𝑎1 (𝑛1 + 1)𝒂2 ]

These vector and matrix have order of

𝑛𝑆 = 𝑛1 𝑛2 + 𝑛1 + 𝑛2
Maximum operation: 𝑇𝑆 = MAX 𝑇1 ,𝑇2
Theorem 4. (Neuts, 1994)
MAX 𝑇1 ,𝑇2 has a PH distribution with representation (𝑛𝑆 , 𝒂𝑆 , 𝑨𝑆 ) given by
𝑛𝑆 = 𝑛1 𝑛2 + 𝑛1 + 𝑛2
𝒂𝑆 = [𝒂1 ⊗ 𝒂2 , 𝑎2 (𝑛2 + 1)𝒂1 , 𝑎1 (𝑛1 + 1)𝒂2 ]
𝑨1 ⊕ 𝑨2 𝑰𝑛1 ⊗ 𝒃2 𝒃1 ⊗ 𝑰𝑛2
𝑨𝑆 = 𝟎 𝑨1 𝟎
𝟎 𝟎 𝑨2
where
𝒃𝑖 = −𝑨𝑖 𝒖
and 𝑎𝑖 (𝑛𝑖 + 1) is the probability of starting at the absorbing state
5. Systems with Basic Structures
Contents
1. System reliability modelling approach
2. Series systems
3. Parallel systems
4. K-out-of-N systems
5. Standby systems
6. Standby systems with imperfect switch
7. Complexity comparison
System reliability modeling approach
• 𝑁 is the number of components in the system
• 𝑇𝑖 , 𝑖 = 1, 2, … , 𝑁, is the lifetime of the 𝑖𝑡ℎ -component
• All components do not undergo instantaneous failure
• 𝑇𝑆 is the system lifetime
• Based on PH distribution closure properties
• Relates the components’ lifetimes to the system lifetime
𝑇𝑆 = 𝜑(𝑇1 , 𝑇2 , . . . , 𝑇𝑁 )
where 𝜑: the system structure function, depends on the system structure.
• Assumed 𝑇𝑖 , 𝑖 = 1, 2, … , 𝑁, are independent non-identical random variables having
PH distributions with representations (𝑛𝑖 , 𝒂𝑖 , 𝑨𝑖 ) with zero jump, i.e. 𝑎𝑖 𝑛𝑖 + 1 = 0
• Closure properties guarantee that the system lifetime also has a PH distribution
Series system: Modeling
• A system that fails if and only if one of its components fails.

• Modeled as
𝑇𝑆 = min( 𝑇1 , 𝑇2 , . . . , 𝑇𝑁 )
• Define
(𝑖)
𝑇𝑆 = min( 𝑇1 , 𝑇2 , . . . , 𝑇𝑖 )
• The system lifetime can be written recursively as
(𝑖) (𝑖−1)
𝑇𝑆 = min( 𝑇𝑆 , 𝑇𝑖 ) for 𝑖 = 2,3, … , 𝑁
(1)
with initial condition 𝑇𝑆 = 𝑇1 is known to have a PH distribution
• Following Theorem 3 recursively shows that 𝑇𝑆 also has PH distribution.
Series system: Lemma
Lemma 1. (Alkaff and Qomarudin, 2020)

𝑇𝑆 of a series system has a PH representation 𝑛𝑆 , 𝒂𝑆 , 𝑨𝑆 where

(𝑖) (𝑖−1)
𝑛𝑆 = 𝑛𝑆 𝑛𝑖
𝑖 𝑖−1
𝒂𝑆 = 𝒂𝑆 ⊗ 𝒂𝑖
𝑖 𝑖−1
𝑨𝑆 = 𝑨𝑆 ⊕ 𝑨𝑖

for 𝑖 = 2, 3, … , 𝑁, with initial conditions

(1) (1) (1)
𝑛𝑆 = 𝑛1 𝒂𝑆 = 𝒂1 𝑨𝑆 = 𝑨1

where (𝑖) superscript denotes the cumulative operation of the first 𝑖 components and 𝑖
subscript denotes the 𝑖𝑡ℎ component.
Series system: Algorithm
• Lemma 1 provides an algorithm for generating a PH representation of 𝑇𝑆
• Presented in MATLAB as function phminsys utilizing the Kronecker function
• The inputs are 𝑁 and (𝒏𝑖 , 𝒂𝑖 , 𝑨𝑖 ) for 𝑖 = 1,2, … , 𝑁
• The output is (𝑛𝑆 , 𝒂𝑆 , 𝑨𝑆 ) of order 𝑛𝑆 = 𝑛𝑁
(𝑖) (𝑖−1)
Script for 𝑇𝑆 = min(𝑇1 , 𝑇2 ) Script for 𝑇𝑆 = min( 𝑇𝑆 , 𝑇𝑖 )
function Ts = phmin(T1,T2) function Ts = phminsys(T)
N = length(T);
Ts.a = kron(T1.a,T2.a);
Ts = T(1);
Ts.A = kronsum(T1.A,T2.A);
for i = 2:N
Ts.n = T1.n*T2.n; T1 = Ts;
end T2 = T(i);
Ts = phmin(T1,T2);
end
end
Series system: Example
• Identical components having 2-stage Erlang distributed lifetime with MTTF = 100
Scalar representation:
𝑅𝑖 𝑡 = 𝑒 −0.02𝑡 + 0.02𝑡𝑒 −0.02𝑡 = 𝐸𝑟(2,100)
𝑅𝑆 𝑡 = 𝐸𝑟(2,100)𝑁
Matrix representation:
−0.02 0.02 1
𝑅𝑖 𝑡 = 1 0 exp 𝑡
0 −0.02 1
−0.02 0.02
𝒂= 1 0 ,𝑨=
0 −0.02
𝑅𝑆 (𝑡) = 𝒂𝑆 exp 𝑨𝑆 𝑡 𝒖𝑆 , where 𝒂𝑆 and 𝑨𝑆
are calculated recursively using Lemma 1
Parallel system: Modeling
• A system that fails if and only if all of its components fail.
• Modeled as
𝑇𝑆 = max( 𝑇1 , 𝑇2 , . . . , 𝑇𝑁 )
• Define
(𝑖)
𝑇𝑆 = max( 𝑇1 , 𝑇2 , . . . , 𝑇𝑖 )
• The system lifetime can be written recursively as
(𝑖) (𝑖−1)
𝑇𝑆 = max( 𝑇𝑆 , 𝑇𝑖 ) for 𝑖 = 2,3, … , 𝑁
with initial condition
(1)
𝑇𝑆 = 𝑇1 is known to have a PH distribution
• Following Theorem 4 recursively shows that 𝑇𝑆 also has PH distribution.
Parallel system: Lemma
Lemma 2. (Alkaff and Qomarudin, 2020)
𝑇𝑆 of a parallel system has a PH representation 𝑛𝑆 , 𝒂𝑆 , 𝑨𝑆 where
(𝑖) (𝑖−1) (𝑖−1)
𝑛𝑆 = 𝑛𝑆 𝑛𝑖 + 𝑛𝑆 + 𝑛𝑖
𝑖 𝑖−1
𝒂𝑆 = 𝒂𝑆 ⊗ 𝒂𝑖 , 𝟎, 𝟎
𝑖−1 𝑖−1
𝑨𝑆 ⊕ 𝑨𝑖 𝑰𝑛 𝑖−1 ⊗ 𝒃𝑖 𝒃𝑆 ⊗ 𝑰𝑛𝑖
𝑆
𝑖
𝑨𝑆 = 𝑖−1
𝟎 𝑨𝑆 𝟎
𝟎 𝟎 𝑨𝑖
for 𝑖 = 2, 3, … , 𝑁, with initial conditions
(1) (1) (1)
𝑛𝑆 = 𝑛1 𝒂𝑆 = 𝒂1 𝑨𝑆 = 𝑨1
where
𝑖−1 𝑖−1
𝒃𝑖 = −𝑨𝑖 𝒖 and 𝒃𝑆 = −𝑨𝑆 𝒖
Parallel system: Algorithm
• Lemma 2 provides an algorithm for generating a PH representation of 𝑇𝑆
• Presented in MATLAB as function phmaxsys utilizing the Kronecker function
• The inputs are 𝑁 and (𝒏𝑖 , 𝒂𝑖 , 𝑨𝑖 ) for 𝑖 = 1,2, … , 𝑁
• The output is (𝑛𝑆 , 𝒂𝑆 , 𝑨𝑆 ) of order 𝑛𝑆 = 𝑛 + 1 𝑁

(𝑖) (𝑖−1)
Script for 𝑇𝑆 = max(𝑇1 , 𝑇2 ) Script for 𝑇𝑆 = max( 𝑇𝑆 , 𝑇𝑖 )
function Ts = phmax(T1,T2) function Ts = phmaxsys(T)
u1 = sparse(ones(T1.n,1)); N = length(T);
u2 = sparse(ones(T2.n,1)); Ts = T(1);
I1 = speye(T1.n); for i = 2:N
I2 = speye(T2.n); T1 = Ts;
z21 = sparse(T1.n,T1.n*T2.n); T2 = T(i);
z31 = sparse(T2.n,T1.n*T2.n); Ts = phmax(T1,T2);
z32 = sparse(T2.n,T1.n); end
z23 = sparse(T1.n,T2.n); end
Ts.a = [kron(T1.a,T2.a), (1-sum(T2.a))*T1.a, (1-sum(T1.a))*T2.a];
Ts.A = [kronsum(T1.A,T2.A) kron(I1,-T2.A*u2) kron(-T1.A*u1,I2);
z21 T1.A z23 ;
z31 z32 T2.A ];
Ts.n = T1.n*T2.n+T1.n+T2.n;
end
Parallel system: Example
• Identical components having 2-stage Erlang distributed lifetime with MTTF = 100
Scalar representation:
𝑅𝑖 𝑡 = 𝑒 −0.02𝑡 + 0.02𝑡𝑒 −0.02𝑡 = 𝐸𝑟(2,100)
𝑅𝑆 𝑡 = 1 − (1 − 𝐸𝑟(2,100))𝑁

Matrix representation:
−0.02 0.02 1
𝑅𝑖 𝑡 = 1 0 exp 𝑡
0 −0.02 1
−0.02 0.02
𝒂= 1 0 ,𝑨=
0 −0.02
𝑅𝑆 (𝑡) = 𝒂𝑆 exp 𝑨𝑆 𝑡 𝒖𝑆 , where 𝒂𝑆 and 𝑨𝑆
are calculated recursively using Lemma 2
K-out-of-N system: Modeling
• A system that operates if and only if at least 𝑲-out-of-𝑵 components are in
acceptable condition
• Modelled as
𝑇𝑆 = kofn( 𝑇1 , 𝑇2 , . . . , 𝑇𝑁 ; 𝐾)
where 𝐾 = 1 represents an 𝑁-parallel system
• PH representation of an 𝑁-parallel system is composed of block matrices associated
with combinations 1 of 𝑁 to 𝑁 of 𝑁 of its components
• PH representation of an 𝐾-out-of-𝑁 system is composed of block matrices
associated with combinations 𝐾 of 𝑁 to 𝑁 of 𝑁 of its components
• Can be obtained by eliminating block matrices associated with combination 1 of 𝑁
to 𝐾 − 1 of 𝑁 from PH representation of 𝑁-parallel system.
• Difficulty occurs due to the locations of the eliminated blocks are scattered
• Need an algorithm to locate their respective rows/columns
K-out-of-N system: Modeling
Obtaining CTMC of a K-out-of-3 system from CTMC of a 3-parallel system
K-out-of-N system: Modeling
• PH representation of an 3-parallel system is composed by block matrices associated
with combination 1 of 3 to 3 of 3 of its 3 components
• PH representation for 2-out-of-3 system is obtained by discarding block matrices
associated with combination 1 of 3 from PH representation of 3-parallel system.
K-out-of-N system: Modeling
Number of Good Row/Column Position 3 out of 4 system
Block Transition Matrix
Components From To
𝑩1 4 𝑨1 ⊕ 𝑨2 ⊕ 𝑨3 ⊕ 𝑨4 𝐿0 + 1 = 1 𝑛1𝑛2𝑛3𝑛4 = 𝐿1
𝑩2 3 𝑨1 ⊕ 𝑨3 ⊕ 𝑨4 𝐿1 + 1 𝐿1 + 𝑛1𝑛3𝑛4 = 𝐿2
𝑩3 3 𝑨2 ⊕ 𝑨3 ⊕ 𝑨4 𝐿2 + 1 𝐿2 + 𝑛2𝑛3𝑛4 = 𝐿3
𝑩4 3 𝑨1 ⊕ 𝑨2 ⊕ 𝑨4 𝐿3 + 1 𝐿3 + 𝑛1𝑛2𝑛4 = 𝐿4
𝑩5 2 𝑨1 ⊕ 𝑨4 𝐿4 + 1 𝐿4 + 𝑛1𝑛4 = 𝐿5
𝑩6 2 𝑨2 ⊕ 𝑨4 𝐿5 + 1 𝐿5 + 𝑛2𝑛4 = 𝐿6
𝑩7 2 𝑨3 ⊕ 𝑨4 𝐿6 + 1 𝐿6 + 𝑛3𝑛4 = 𝐿7
𝑩8 3 𝑨1 ⊕ 𝑨2 ⊕ 𝑨3 𝐿7 + 1 𝐿7 + 𝑛1𝑛2𝑛3 = 𝐿8 4-parallel
𝑩9 2 𝑨1 ⊕ 𝑨3 𝐿8 + 1 𝐿8 + 𝑛1𝑛3 = 𝐿9
𝑩10 2 𝑨2 ⊕ 𝑨3 𝐿9 + 1 𝐿9 + 𝑛2𝑛3 = 𝐿10
𝑩11 2 𝑨1 ⊕ 𝑨2 𝐿10 + 1 𝐿10 + 𝑛1𝑛2 = 𝐿11 Discarded
𝑩12 1 𝑨1 𝐿11 + 1 𝐿11 + 𝑛1 = 𝐿12 block
𝑩13 1 𝑨2 𝐿12 + 1 𝐿12 + 𝑛2 = 𝐿13 Matrices for
𝑩14 1 𝑨3 𝐿13 + 1 𝐿13 + 𝑛3 = 𝐿14 3-out-of-4
𝑩15 1 𝑨4 𝐿14 + 1 𝐿14 + 𝑛4 = 𝐿15
𝑩16 0 0 𝐿15 + 1 𝐿15 + 1
K-out-of-N system: Lemma
Lemma 3. (Alkaff and Qomarudin, 2020)

𝑇𝑆 of 𝐾-out-of-𝑁 systems has a PH representation (𝑛𝑆 , 𝒂𝑆 , 𝑨𝑆 ) obtained by discarding

the rows and columns of an infinitesimal generator 𝑸𝑆 , and the columns of the initial
state probability vector 𝒂𝑆 of the 𝑁-parallel system that represent less than 𝐾 good
components identified using the following algorithms.

Algorithm for identifying the number of Algorithm for calculating the dimension of
good components in 𝑸𝑆 each block matrices in 𝑸𝑆
𝒌 = [1] 𝒏𝑩 = [𝑛1 ]
for 𝑗 = 2 to 𝑁 for 𝑗 = 2 to 𝑁
𝒌 = 𝒌 + 𝒖𝑇 , 𝒌, 1 ; 𝒏𝑩 = [𝒏𝑩 𝑛𝑗 , 𝒏𝑩 , 𝑛𝑗 ];
end end
𝒌 = [𝒌, 0] 𝒏𝑩 = [𝒏𝑩 , 1]
K-out-of-N system: Lemma
• The first recursion identifies the quantity of acceptable components in each block:
𝑁 = 2: 𝒌 = [2, 1, 1, 0]
𝑁 = 3: 𝒌 = [ 2, 1, 1 + [1, 1, 1], 2, 1, 1, 1, 0] = [3, 2, 2, 2, 1, 1, 1, 0]
𝑁 = 4: 𝒌 = [4, 3, 3, 3, 2, 2, 2, 3, 2, 2, 2, 1, 1, 1, 1, 0]
𝑁 = 5: 𝒌 = [5, 4, 4, 4, 3, 3, 3, 4, 3, 3, 3, 2, 2, 2, 2, 4, 3, 3, 3, 2, 2, 2, 3, 2, 2, 2, 1, 1, 1, 1, 1, 0]
• 𝑨𝑆 of a 3-out-of-5 system is block in 𝑸𝑆 of a 5-parallel system without 𝑘𝑗 = 2,1,0
• The second recursion is to identify the size of each block
𝑁 = 2: 𝒏𝑩 = [𝑛1 𝑛2 , 𝑛1 , 𝑛2 , 1]
𝑁 = 3: 𝒏𝑩 = 𝑛1 𝑛2 , 𝑛1 , 𝑛2 𝑛3 , (𝑛1 𝑛2 , 𝑛1 , 𝑛2 ), 𝑛3 , 1 = 𝑛1 𝑛2 𝑛3 , 𝑛1 𝑛3 , 𝑛2 𝑛3 , 𝑛1 𝑛2 , 𝑛1 , 𝑛2 , 𝑛3 , 1
𝑁 = 4: 𝒏𝑩 =
[ 𝑛1 𝑛2 𝑛3 , 𝑛1 𝑛3 , 𝑛2 𝑛3 , 𝑛1 𝑛2 , 𝑛1 , 𝑛2 , 𝑛3 𝑛4 , (𝑛1 𝑛2 𝑛3 , 𝑛1 𝑛3 , 𝑛2 𝑛3 , 𝑛1 𝑛2 , 𝑛1 , 𝑛2 , 𝑛3 ), 𝑛4 , 1]
𝑗 𝑗−1 𝑗
𝐿𝑗 = ∑𝑖=1 𝑛𝑩𝑖 → first and last rows/columns of 𝑩𝑗 is ∑𝑖=1 𝑛𝑩𝑖 + 1 and ∑𝑖=1 𝑛𝑩𝑖
• For 3-out-of-5 system, that must be discarded are row/column number ൫∑11 𝑖=1 𝑛𝑩𝑖 +
1൯ to ∑15
𝑖=1 𝑛𝑩 𝑖
, ∑ 19
𝑖=1 𝑛𝑩 𝑖
+ 1 to ∑ 22
𝑖=1 𝑛 𝑩 𝑖
, and ∑ 23
𝑖=1 𝑛𝑩 𝑖
+ 1 to ∑ 32
𝑖=1 𝑛𝑩𝑖
K-out-of-N system: Algorithm
• Lemma 3 is presented in MATLAB as function phkofnsys
• The inputs are 𝐾, 𝑁 and (𝒏𝑖 , 𝒂𝑖 , 𝑨𝑖 ) for 𝑖 = 1,2, … , 𝑁
• The output is (𝑛𝑆 , 𝒂𝑆 , 𝑨𝑆 ) of order 𝑛𝑆 = 𝑛 + 1 𝑁 − 𝑛+1 𝐾−1

Script for 𝑇𝑆 = kofn( 𝑇1 , 𝑇2 , . . . , 𝑇𝑁 ; 𝐾) (based on phmaxsys)

function Ts = phkofnsys(T,K) .....
N = length(T); for j = 1:K-1
Ts = phmaxsys(T); i = 1;
Q = 1; while i <= M
L = T(1).n; if Q(i) == j
P(1) = 1;
for i = 2:N for ii = 2:length(L)+1
Q = [Q+1 Q 1]; P(ii) = P(ii-1)+L(ii-1);
end end
Ts.A(:,P(i):P(i+1)-1) = [];
for i = 2:N Ts.A(P(i):P(i+1)-1,:) = [];
L = [L*T(i).n L T(i).n]; Ts.a(P(i):P(i+1)-1) = [];
end Q(i) = []; L(i) = [];
M = M-1; i = i-1;
M = length(Q); clear P
end
..... i = i+1;
end
end
Ts.n = length(Ts.a);
end
K-out-of-N system: Example
• Identical components having 2-stage Erlang distributed lifetime with MTTF = 100
Scalar representation:
𝑅𝑖 𝑡 = 𝑒 −0.02𝑡 + 0.02𝑡𝑒 −0.02𝑡 = 𝐸𝑟(2,100)
𝑁

𝑅𝑆 (𝑡) = ෍ 𝐶 𝑗, 𝑁 𝐸𝑟(2,100)𝑗 (1 − 𝐸𝑟(2,100))𝑁−𝑗

𝑗=𝐾

Matrix representation:
−0.02 0.02 1
𝑅𝑖 𝑡 = 1 0 exp 𝑡
0 −0.02 1
−0.02 0.02
𝒂𝑖 = 1 0 , 𝑨𝑖 =
0 −0.02
𝑅𝑆 (𝑡) = 𝒂𝑆 exp 𝑨𝑆 𝑡 𝒖𝑆 , where 𝒂𝑆 and 𝑨𝑆
are calculated recursively using Lemma 3
Standby system: Modeling
• A structure of redundancy in which the standby component is called upon when
the operating component failed
• Hence,
𝑇𝑆 = 𝑇1 + 𝑇2 + ⋯ + 𝑇𝑁 = add( 𝑇1 , 𝑇2 , . . . , 𝑇𝑁 )
• Define
(𝑖)
𝑇𝑆 = add( 𝑇1 , 𝑇2 , . . . , 𝑇𝑖 )
• The system lifetime can be written recursively as
(𝑖) (𝑖−1)
𝑇𝑆 = add( 𝑇𝑆 , 𝑇𝑖 ) for 𝑖 = 2,3, … , 𝑁
with initial condition
(1)
𝑇𝑆 = 𝑇1 is known to have a PH distribution
• Following Theorem 2, recursively shows that 𝑇𝑆 also has PH distribution.
Standby system: Lemma
Lemma 4. (Alkaff and Qomarudin, 2020)
𝑇𝑆 of a standby system with a perfect switch has a PH representation 𝑛𝑆 , 𝒂𝑆 , 𝑨𝑆 where
(𝑖) (𝑖−1)
𝑛𝑆 = 𝑛𝑆 + 𝑛𝑖
𝑖 𝑖−1
𝒂𝑆 = 𝒂𝑆 ,𝟎
𝑖−1 𝑖−1
𝑖 𝑨𝑆 𝒃𝑆 𝒂𝑖
𝑨𝑆 =
𝟎 𝑨𝑖

for 𝑖 = 2, 3, … , 𝑁, with initial conditions

(1) (1) (1)
𝑛𝑆 = 𝑛1 𝒂𝑆 = 𝒂1 𝑨𝑆 = 𝑨1

and
𝑖−1 𝑖−1
𝒃𝑆 = −𝑨𝑆 𝒖
Standby system with imperfect switch: Modeling
• In a two-component standby system with imperfect switch, the system lifetime is
given by
𝑇𝑆 = 𝑇1 if the switch is not functioning (with probability 1 − 𝑝1 )
𝑇𝑆 = 𝑇1 + 𝑇2 if the switch is functioning (with probability 𝑝1 )
• Can be written as
𝑇𝑆 = 𝑇1 + 𝑝1 𝑇2
• Consequently,
𝒂𝑆 = [𝒂1 , 𝟎]
𝑨1 𝑝1 𝒃1 𝒂2
𝑨𝑆 =
𝟎 𝑨2

• If 𝑝1 = 1, then it is reduced to ordinary standby system

Standby system with imperfect switch: Modeling
• A structure of standby redundancy connected using imperfect switch with
probability 𝑝𝑖 functioning properly when component 𝑖 is failed
• Can be modelled as,
𝑇𝑆 = add(𝑇1 , 𝑝1 , 𝑇2 , 𝑝2 , . . . , 𝑝𝑁−1 , 𝑇𝑁 )
• Define
(𝑖)
𝑇𝑆 = add( 𝑇1 , 𝑝1 , 𝑇2 , 𝑝2 , . . . , 𝑝𝑖−1 , 𝑇𝑖 )
• The system lifetime can be written recursively as
(𝑖) 𝑖−1
𝑇𝑆 = add( 𝑇𝑆 , 𝑝𝑖−1 , 𝑇𝑖 ) for 𝑖 = 2,3, … , 𝑁
with initial condition
(1)
𝑇𝑆 = 𝑇1 is known to have a PH distribution
• Following Theorem 2 recursively shows that 𝑇𝑆 also has PH distribution.
Standby system with imperfect switch: Lemma
Lemma 5. (Alkaff and Qomarudin, 2020)
𝑇𝑆 of a standby system with a switching probability 𝑝𝑖 that the switch function properly upon
the failure of component 𝑖 has a PH representation 𝑛𝑆 , 𝒂𝑆 , 𝑨𝑆 where
(𝑖) (𝑖−1)
𝑛𝑆 = 𝑛𝑆 + 𝑛𝑖
𝑖 𝑖−1
𝒂𝑆 = 𝒂𝑆 ,𝟎
(𝑖−1) 𝑖−1
(𝑖) 𝑨 𝒃𝑆 𝒂𝑖
𝑨𝑆 = 𝑆
𝟎 𝑨𝑖
for 𝑖 = 2, 3, … , 𝑁, with initial conditions
(1) (1) (1)
𝑛𝑆 = 𝑛1 𝒂𝑆 = 𝒂1 𝑨𝑆 = 𝑨1
where
𝑖−1 𝟎
𝒃𝑆 𝒂𝑖 = and 𝒃𝑖−1 = −𝑨𝑖−1 𝒖
𝑝𝑖−1 𝒃𝑖−1 𝒂𝑖
*Note that Lemma 5 is reduced to Lemma 4 for 𝑝 = 1 → need one algorithm for both
Standby system with imperfect switch: Algorithm
• Lemma 5 is presented in MATLAB as function phaddsys
• Applicable for both perfect and imperfect switches
• The inputs: 𝑁, (𝒏𝑖 , 𝒂𝑖 , 𝑨𝑖 ) for 𝑖 = 1,2, … , 𝑁, and 𝑝𝑖 for 𝑖 = 1, … , 𝑁 − 1
• The output is (𝑛𝑆 , 𝒂𝑆 , 𝑨𝑆 ) of order 𝑛𝑆 = 𝑛𝑁
Script for 𝑇𝑆 = add( 𝑇1 , 𝑇2 , . . . , 𝑇𝑁 ) Script for 𝑇𝑆 = add( 𝑇1 , 𝑝1 , 𝑇2 , 𝑝2 , . . . , 𝑝𝑁−1 , 𝑇𝑁 )
function Ts = phadd(T1,T2,T3,p) function Ts = phaddsys(T,P)
u2 = sparse(ones(T2.n,1)); N = length(T);
z31 = sparse(T3.n,T1.n); Ts = T(1);
Ts.a = [T1.a, (1-sum(T1.a))*T3.a];
for i = 2:N
if (T1.n-T2.n) == 0 T1 = Ts;
a13 = -p*T2.A*u2*T3.a; T2 = T(i-1);
else T3 = T(i);
a13 = [sparse(T1.n-T2.n,T3.n); Ts = phadd(T1,T2,T3,P(i-1));
-p*T2.A*u2*T3.a]; end
end
end
Ts.A = [T1.A a13 ;
z31 T3.A ];
Ts.n = length(Ts.a);
end
Standby system with perfect switch: Example
• Identical components having 2-stage Erlang distributed lifetime with MTTF = 100
Scalar representation:
𝑅𝑖 𝑡 = 𝑒 −0.02𝑡 + 0.02𝑡𝑒 −0.02𝑡 = 𝐸𝑟(2,100)
𝑅𝑆 𝑡 = 𝑅1 𝑡 ∗ 𝑅2 𝑡 ∗ ⋯ ∗ 𝑅𝑁 𝑡
= 𝐸𝑟(2𝑁, 100)
(*: convolution)
Matrix representation:
−0.02 0.02 1
𝑅𝑖 𝑡 = 1 0 exp 𝑡
0 −0.02 1
−0.02 0.02
𝒂𝑖 = 1 0 , 𝑨𝑖 =
0 −0.02
𝑅𝑆 (𝑡) = 𝒂𝑆 exp 𝑨𝑆 𝑡 𝒖𝑆 , where 𝒂𝑆 and 𝑨𝑆
are calculated recursively using Lemma 4
Standby system with imperfect switch: Example
• Identical components having 2-stage Erlang distributed lifetime with MTTF = 100
Scalar representation for 𝑝 = 0.9, 𝑁 = 8:
𝑅𝑖 𝑡 = 𝑒 −0.02𝑡 + 0.02𝑡𝑒 −0.02𝑡 = 𝐸𝑟(2,100)
𝑅𝑆 𝑡 = 0.1𝐸𝑟(2,100) + 0.9⨉0.1𝐸𝑟(4,100)
+ ⋯ + 0.96 ⨉0.1𝐸𝑟(2 𝑁 − 1 , 100)
+0.97 𝐸𝑟(2𝑁, 100)
Matrix representation:
−0.02 0.02 1
𝑅𝑖 𝑡 = 1 0 exp 𝑡
0 −0.02 1
−0.02 0.02
𝒂𝑖 = 1 0 , 𝑨𝑖 =
0 −0.02
𝑅𝑆 (𝑡) = 𝒂𝑆 exp 𝑨𝑆 𝑡 𝒖𝑆 , where 𝒂𝑆 and 𝑨𝑆
are calculated recursively using Lemma 5
Complexity comparison

Complexity
No Structure PH Representation Scalar Representation
(Matrix Dimension) (Number of Terms)
1 Series 𝑛𝑁 𝑛𝑁
2 Parallel 𝑛+1 𝑁 𝑛+1 𝑁
3 K-out-of-N 𝑛 + 1 𝑁 − 𝑛 + 1 𝐾−1 𝑛 + 1 𝑁 − 𝑛 + 1 𝐾−1
4 Standby with perfect switch 𝑛𝑁 𝑛𝑁
5 Standby with imperfect switch 𝑛𝑁 𝑛𝑁+1
6. Systems with General Structures
Contents
1. Modeling system reliability
2. Algorithm for system reliability analysis
3. Calculating systems reliability measures
4. Calculating confidence interval
Modeling system reliability
• A real system usually comprises many subsystems with different structures.
• A simple coherent system can be decomposed into subsystems with basic structures:
• Series
• Parallel
• K-out-of-N
• Standby with perfect switch
• Standby with imperfect switch
• Let the system:
• can be decomposed into 𝑴 subsystems whose lifetimes are denoted by 𝑆1 , 𝑆2 , … , 𝑆𝑀
• each subsystem 𝑚 can be decomposed into 𝑲𝒎 sub-subsystems whose lifetimes are
denoted by 𝑉𝑚1 , 𝑉𝑚2 , … , 𝑉𝑚𝐾𝑚
• each sub-subsystem 𝑽𝒎𝒌 has 𝑳𝒎𝒌 components whose lifetimes are denoted by
𝑇𝑚,𝑘1 , 𝑇𝑚,𝑘2 , . . . , 𝑇𝑚,𝑘𝐿𝑚𝑘 .
Modeling system reliability
• The system can be expressed mathematically as
𝑉𝑚,𝑘 = 𝜑𝑚,𝑘 (𝑇𝑚,𝑘1 , 𝑇𝑚,𝑘2 , . . . , 𝑇𝑚,𝑘𝐿𝑚𝑘 )
𝑆𝑚 = 𝜑𝑚 𝑉𝑚1 , 𝑉𝑚2 , … , 𝑉𝑚𝐾𝑚
𝑇𝑆 = 𝜑 𝑆1 , 𝑆2 , … , 𝑆𝑀
where 𝜑, 𝜑𝑚 , 𝑎𝑛𝑑 𝜑𝑚,𝑘 are structure functions, and each has a basic structure.
• 𝑇𝑚,𝑘𝑙𝑚𝑘 has a PH distribution, hence 𝑉𝑚,𝑘 , 𝑆𝑚 , 𝑇𝑆 must has PH distributions.
• 𝑇𝑆 can be generated hierarchically using the recursion algorithms:
component level → sub-subsystem level → subsystem → system level.
System

Subsystem ... Subsystem

Sub-subsystem ... Sub-subystem

Component ... Component

Algorithm for system reliability analysis: example

Algorithm
𝑉21 = min 𝑇4 , 𝑇5 , 𝑇6
𝑉22 = min 𝑇7 , 𝑇8
𝑆1 = add 𝑇1 , 𝑝, 𝑇2 , 𝑝, 𝑇3
𝑆2 = max 𝑉21 , 𝑉22
𝑆3 = kofn 𝑇9 , 𝑇10 , 𝑇11 , 2
𝑇𝑠 = min 𝑆1 , 𝑆2 , 𝑆3

Main Program (MATLAB)

V21 = phminsys([T(4),T(5),T(6)]);
• S1 is a 3-standby with imperfect switch (using add) V22 = phminsys([T(7),T(8)]);
• S2 is a parallel-series: S1 = phaddsys([T(1),T(2),T(3)],p);
S2 = phmaxsys([V21,V22]);
• 3-series V21 (using min) S3 = phkofnsys([T(9),T(10),T(11)],2);
• 2-series V22 (using min) Paralleled (using max)
Ts = phminsys([S1,S2,S3]);
• S3 is a 2-out-of3 (using kofn)
Calculating system reliability measures
• Algorithm for calculating system’s hazard function from 𝑅𝑆 (𝑡) = 𝒂𝑆 𝑒 𝑨𝑆𝑡 𝒖𝑆

t = linspace(0,t_max,samples); % vector of time

for i = 1:samples
e = expmq(Ts.A*t(i));
Rs(i) = sum(Ts.a*e); % vector of Rs values
hs(i) = -sum(Ts.a*Ts.A*e)/Rs(i); % vector of hs values
end

• Algorithm for calculating MTTF and SDTTF from 𝑅𝑆 (𝑡) = 𝒂𝑆 𝑒 𝑨𝑆𝑡 𝒖𝑆

mttf = -sum(Ts.a/Ts.A);
vttf = 2*sum(Ts.a/(Ts.A^2))-mttf^2;
sdttf = sqrt(vttf);
A system example
• The system has components with iid lifetimes following PH distributions:
No Distribution Type CVTTF MTTF PH Representation
1 Exponential (𝜆) CFR 1 100 𝑛=1 𝒂= 1 𝑨= −𝜆
−𝑘𝜆 𝑘𝜆 0 0
1/√𝑘 0 −𝑘𝜆 𝑘𝜆 0
2 𝑘-stage Erlang IFR 100 𝑛 = 𝑘 𝒂 = 1,0, … , 0 𝑨 =
(< 1) ⋮ ⋮ ⋱ ⋮
0 0 0 −𝑘𝜆
−𝜆1 𝑝𝜆1
3 2-stage Coxian DFR >1 100 𝑛=2 𝒂= 1 0 𝑨=
0 −𝜆2
−0.02 0.018
4 2-stage CPH* IFR 0.7976 100 𝑛=2 𝒂= 1 0 𝑨 = 0.796703
0.004 −0.04
* CPH: Cyclic PH distribution
Generating systems reliability measures
• Plotting system’s reliability and hazard functions
−1
𝑅𝑆 𝑡 = 𝒂𝑆 𝑒 𝑨𝑆𝑡 𝒖𝑆 ℎ𝑆 𝑡 = −𝒂𝑆 𝑨𝑆 𝑒 𝑨𝑆𝑡 𝒖𝑆 𝒂𝑆 𝑒 𝑨𝑆𝑡 𝒖𝑆

2-Erlang

2-Erlang
CPH

CPH Exponential

Exponential 2-Coxian

2-Coxian
Generating systems reliability measures
• Calculated and plotted system’s MTTF and SDTTF
Components System
Distribution MTTF CVTTF Type MTTF SDTTF
Exponential 100 1.0000 CFR 42.0418 29.0885
2- Erlang 100 0.7071 IFR 58.6044 26.3733
3- Erlang 100 0.5774 IFR 66.3844 23.6086
4- Erlang 100 0.5000 IFR 71.0564 21.4834
5- Erlang 100 0.4472 IFR 74.2419 19.8336
2- Coxian 100 1.5000 DFR 31.8923 23.5001
2 Coxian 100 2.0000 DFR 27.3290 19.9697
Cyclic PH 100 0.7976 IFR 53.9841 27.0343

* The system MTTF using CPH fits precicely into graph of MTTF using Erlang and Coxian
Generating system reliability measures: CI for MTTF
• Let 𝑇𝑆1 , 𝑇𝑆2 , … , 𝑇𝑆𝑀 be the system lifetimes obtained from the 𝑴 measurements.
The average of the system TTF calculated from these sample data is
1 𝑆𝑇
ഥ
𝑇𝑆 = 𝑇𝑆1 + 𝑇𝑆2 + ⋯ + 𝑇𝑆𝑀 =
𝑀 𝑀
• 𝑆𝑇 is a PH distributed whose representation can be obtained using phadd
• The exact (1 − 𝛼) CI for the system MTTF can be obtained from 𝐹𝑆𝑇 (𝑡) as the values
of 𝒕/𝑴 for which 𝐹𝑆𝑇 𝑡 = 1 − 𝛼/2 and 𝐹𝑆𝑇 𝑡 = 𝛼/2.
Components System
Distribution MTTF M MTTF 95% CI for MTTF
10 42.0418 26.1538 61.5385
15 42.0418 28.2051 57.8205
Exponential 100 20 42.0418 30.0000 55.6250
25 42.0418 31.3846 54.1538
30 42.0418 32.2756 52.9808
• May use normal approximation using system’s MTTF and SDTTF values
7. Discussion
Discussion
• Functional system reliability analysis is made possible by utilizing some closure
properties of PH distribution.
• The result is algorithms readily programmed in MATLAB, available at (including all
examples and comparison with symbolic processing):
https://github.com/mochamadnurq/reliability-basicsystem-phasetype
• The method is exact and applicable for systems with general structures that may
contains DFR, IFR, CFR, and even BFR components.
• Possible extensions of the model are to include:
• dependent components
• warm (or cold with deterioration) standbys
• multi-state components
• bridge structure (submitted for publication)
• application for binary decision diagram (BDD) and phased mission system
• Challenge is reducing computation time by:
• exploiting the resulted matrix structures
• using numerical techniques specifically for large matrices with Kronecker structures.